1. CHAN Siu Lung, Daniel CHAN Wai Kin, Ken CHOW Chin Hung, Victor KOON Ping Yin, Bob SPRINT: A Scalable Parallel Classifier for Data Mining

2. Overview Classification Decision Tree SPRINT Serial Algorithm Parallelizing Classification Performance Evaluation Conclusion Reference

3. Classification Objective –To build a model of the classifying attribute based upon the other attributes Classification given –A set of example record, called a training set that is a set of records consists of several attributes Attributes are either 1.Continuous (i.e. Age: 23,24,25) 2.Categorical (i.e. Gender: Female, Male) One of the attributes called the classifying attribute

4. Decision Trees Advantages: –Good for data mining –Can be constructed fast –Models are simple –Easy to understand –Can be easily converted into SQL statements –Classifiers obtain better accuracy

5. Example – Car Insurance Continuous attribute Categorical attribute

6. Classification algorithm Two classification algorithms are presented in the paper –SLIQ –SPRINT SLIQ –Handles large disk-resident data –Requires some data per record that stay in memory-resident all the time SPRINT –Removes all of the memory restrictions –Fast and scalable and it can be easily parallelized How good it is? –Excellent scaleup, speedup and sizeup properties

7. SPRINT Serial Algorithm A decision tree classifier is built in two phases –Growth phase –Prune phase Growth phase is computationally much more expensive than pruning

8. Serial Algorithm Growth Phase Key Issue –To find split points that define node tests. –Having chosen a split point, how to partition the data Data Structures –Attribute lists –Histograms

9. Serial Algorithm Data Structure (Attribute lists) SPRINT creates an attribute list for each attribute Entries are called attribute records which contains –Attribute value –Class label –Index of the record

10. Serial Algorithm Data Structure (Attribute lists) The initial lists are associated with the root As the tree is grown, the attribute lists belonging to each node are partitioned and associated with the children

11. Serial Algorithm Data Structure (Histograms) For continuous attributes, two histograms are associated with each decision-tree node. These histograms, denoted as C above and C below –C below: : maintains this distribution for attribute records that already been processed –C above : maintains this distribution for attribute records that have not been processed

12. Serial Algorithm Data Structure (Histograms) For categorical attributes, one histogram associated with a node. However, only one histogram is needed and called count matrix

13. Serial Algorithm Finding split points While growing the tree, the goal at each node is to determine the split point that “best” divides the training records belonging to the leaf In the paper, they use gini index Gini(S)=1-  p j 2 –Where p j is the relative of class j in S Gini split (S) = n 1 /n(S 1 )+n 2 /n(S 2 )

14. Serial Algorithm Example – Split point for the continuous attributes AgeClassTid 17High1 20High5 23High0 32Low4 43High2 68Low3 HL C above 30 C below 12 Cursor Position 3:

15. Serial Algorithm Example – Split point for the continuous attributes After finding all the gini indexes we choose the lowest as the split point Therefore, we split at position 3 where the candidate split point is the mid-point between age 23 and 32 (i.e. Age < 27.5)

16. Serial Algorithm Example – Split point for the categorical attributes HL Family21 Sports20 Truck01

17. Serial Algorithm Performing the split Once the best split point has been found for a node, we then execute the split by creating child nodes and dividing the attribute records between them For the rest of the attribute lists (i.e. CarType) we need to retrieve the information by using rids

18. Serial Algorithm Comparison with SLIQ SLIQ does not have separate sets of attribute lists for each node Advantage –Do not have to rewrite these lists during a split –Reassignment of records is easy Disadvantage –It must stay in memory all the time which limits the amount of data that can be classified by SLIQ SPRINT was not to outperform SLIQ. Rather, the purpose of the algorithm is to develop an accurate classifier for datasets, and to develop a classifier efficiently. Furthermore, SPRINT is designed to be easily parallelizable, thus the workload can be shared among N processor

19. SPRINT Parallelizing Classification SPRINT was specifically designed to remove any dependence on data structures that are either centralized or memory-resident These algorithms all based on a shared-nothing parallel environment where each of N processor has private memory and disks. The processor are connected by a communication network and can communicate only by passing message

20. Parallelizing Classification Data placement and Workload Balancing The main data structures are –attribute lists –class histograms SPRINT achieves uniform data placement and workload balancing by distributing the attribute lists evenly over N processor Each processor to work on only 1/N of the total data

21. Parallelizing Classification Finding split points (Continuous attributes) In a parallel environment, each processor has a separate contiguous section of a “global” attribute list –C below: : must initially reflect the class distribution of all sections of an attribute-list assigned to processors of lower rank –C above : must initially reflect the class distribution of the local section as well as all sections assigned to processor of higher rank

22. Parallelizing Classification Example: Split point for the continuous attributes AgeClassrid 17High1 20High5 23High0 HL C below 00 C above 42 AgeClassrid 32Low4 43High2 68Low3 HL C below 30 C above 12 Processor 1 Processor 0

23. Parallelizing Classification Finding split points (Categorical attributes) Each processor built a counter matrix for a leaf Exchange these matrices to get the “global” counts Sums the local matrices to get the global count matrices

24. Parallelizing Classification Performing the Splits Splitting the attribute lists for each leaf is identical to the serial algorithm The only difference is that before building the probe structure, we will need to collect rids from all the processors. Exchange the rids Each processor constructs a probe-structure with all the rids and using it to split the leaf’s remaining attribute lists

25. Parallelizing Classification Parallelizing SLIQ Two approaches for parallelizing SLIQ –The class list is replicated in memory of every processor (SLIQ/R) –The class list is distributed such that each processor’s memory holds only a portion of the entire list (SLIQ/D) Disadvantage –SLIQ/R – the size of the training set is limited by the memory size of a single processor –SLIQ/D – High communication cost

26. Performance Evaluation The primary metric for evaluating classifier performance –Classification accuracy –Classification time –Size of the decision tree The accuracy and tree size characteristics of SPRINT are identical to SLIQ Focus only on the classification time

27. Performance Evaluation Datasets Each record in synthetic database consists of nine attributes. In the paper, they present results for two of these function Function 2 results in fairly small decision tree Function 7 produces very large tree Both these functions divide the database into two classes

28. Performance Evaluation Serial Performance We used training sets ranging in size from 10,000 records to 2.5 million records To examine how well SPRINT performs in operating regions where SLIQ can and cannot run

29. Performance Evaluation Comparison of Parallel Algorithms In this experiments, each processor contained 50,000 training examples and the number of processors varied from 2 to 16

30. Performance Evaluation Scaleup

31. Performance Evaluation Speedup

32. Performance Evaluation Sizeup

33. Performance Evaluation Conclusion Classification can handle very large database by building classifiers SPRINT efficiently allows classification of virtually any sized dataset and parallelizable. SPRINT scales nicely with the size of the dataset SPRINT parallelize efficiently in a shared- nothing environment

34. References John Shafer,Rakeeh Agrawal and Manish Mehta. SPRINT: A Scalable Parallel Classifier for Data Mining